介绍： Polling and prediction: Epic fail
How a mid-sized error led to a rash of bad forecasts
From The Economist 20161112

Audio:

0:00




AS POLLING errors go, this year’s misfire was not particularly large—at least in the national surveys. Mrs Clinton is expected to win the popular vote by a bit over one perc...

介绍： Polling and prediction: Epic fail
How a mid-sized error led to a rash of bad forecasts
From The Economist 20161112

Audio:

0:00




AS POLLING errors go, this year’s misfire was not particularly large—at least in the national surveys. Mrs Clinton is expected to win the popular vote by a bit over one percentage point once all the ballots are counted, two points short of her projection. That represents a better prediction than in 2012, when Barack Obama beat his polls by three. But America does not choose its president by popular vote, and three of Donald Trump’s bigger outperformances occurred in states around the Great Lakes that proved decisive. Mrs Clinton led the polls in Wisconsin by five points, and in Michigan and Pennsylvania by four; Mr Trump is projected to claim them all, albeit by narrow margins. He did even better in Ohio, where he turned a two-point poll lead into an 8.5-point romp, and Iowa, where a three-point edge became a 9.5-point blowout.

While pollsters correctly gauged the sentiment of most slices of the electorate, they underestimated Mr Trump’s appeal to working-class whites. Although it was clear that he would run up the score with these voters, he managed to exceed even pollsters’ rosy expectations for him: projected to win them by 30 points, the national exit poll showed him winning by 39, a larger edge than Mrs Clinton’s among Latinos. The share of a state’s electorate represented by whites lacking a college degree was an almost perfect predictor of how he did relative to polling (see chart).

It is possible that “shy Trump” voters didn’t want to admit their support to pollsters. However, there was no evidence of such a pattern during the Republican primaries, when Mr Trump did not generally beat his polls. And given his margin with working-class whites, it is hard to imagine that people whose friends and neighbours mainly backed him would be ashamed to say so themselves. A likelier cause is “non-response bias”—that working-class whites who backed Mr Trump were particularly reluctant to answer the phone. It is also possible that some decided to vote Republican after the last polls were completed. Lastly, Mr Trump’s blunt, targeted courtship of this demographic group, which historically has shown a fairly low propensity to vote, may have motivated them to turn out in greater numbers. Such enthusiasm is hard for pollsters to detect.


Whatever the cause, this miss was within the range of reasonable expectations, given that the margin of error is magnified when dealing with demographic subgroups. The key question for forecasters was how a midsized polling mistake led them to get the election so wrong. For models based on state polls, the core issue was how well an error in one state was likely to foreshadow one in the same direction elsewhere—and if so, where. Mr Trump’s six-point outperformance in Wisconsin had little bearing on his performance in Colorado, but spelled doom for Mrs Clinton in nearby Michigan, Ohio and Pennsylvania. Prediction models that either used weaker or less precisely targeted correlations between states were more bullish on her odds, and performed worse.

There is one family of forecasts that did better: those which ignore both polls and candidates and predict results based exclusively on structural factors like economic performance and incumbency. This approach suggested all along that the 2016 campaign was likely to be an extremely tight race. Yet because these models seemed unsophisticated, and because Mr Trump’s campaign was so unusual, they were largely overlooked.

Copyright © 2015-2016 Share2China.com | Powered by Flask and MongoDB